lustre/ptlrpc/sec_bulk.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/sec_bulk.c
  37  *
  38  * Author: Eric Mei <ericm@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_SEC
  42
  43 #include <libcfs/libcfs.h>
  44
  45 #include <obd.h>
  46 #include <obd_cksum.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49 #include <lustre_net.h>
  50 #include <lustre_import.h>
  51 #include <lustre_dlm.h>
  52 #include <lustre_sec.h>
  53
  54 #include "ptlrpc_internal.h"
  55
  56 static int mult = 20 - PAGE_CACHE_SHIFT;
  57 static int enc_pool_max_memory_mb;
  58 CFS_MODULE_PARM(enc_pool_max_memory_mb, "i", int, 0644,
  59                 "Encoding pool max memory (MB), 1/8 of total physical memory by default");
  60
  61
  62 /****************************************
  63  * bulk encryption page pools           *
  64  ****************************************/
  65
  66
  67 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  68 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  69
  70 #define IDLE_IDX_MAX            (100)
  71 #define IDLE_IDX_WEIGHT         (3)
  72
  73 #define CACHE_QUIESCENT_PERIOD  (20)
  74
  75 static struct ptlrpc_enc_page_pool {
  76         /*
  77          * constants
  78          */
  79         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  80         unsigned int     epp_max_pools;   /* number of pools, const */
  81
  82         /*
  83          * wait queue in case of not enough free pages.
  84          */
  85         wait_queue_head_t    epp_waitq;       /* waiting threads */
  86         unsigned int     epp_waitqlen;    /* wait queue length */
  87         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  88         unsigned int     epp_growing:1;   /* during adding pages */
  89
  90         /*
  91          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  92          * this is counted based on each time when getting pages from
  93          * the pools, not based on time. which means in case that system
  94          * is idled for a while but the idle_idx might still be low if no
  95          * activities happened in the pools.
  96          */
  97         unsigned long    epp_idle_idx;
  98
  99         /* last shrink time due to mem tight */
 100         long             epp_last_shrink;
 101         long             epp_last_access;
 102
 103         /*
 104          * in-pool pages bookkeeping
 105          */
 106         spinlock_t       epp_lock;         /* protect following fields */
 107         unsigned long    epp_total_pages; /* total pages in pools */
 108         unsigned long    epp_free_pages;  /* current pages available */
 109
 110         /*
 111          * statistics
 112          */
 113         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 114         unsigned int     epp_st_grows;          /* # of grows */
 115         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 116         unsigned int     epp_st_shrinks;        /* # of shrinks */
 117         unsigned long    epp_st_access;         /* # of access */
 118         unsigned long    epp_st_missings;       /* # of cache missing */
 119         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 120         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 121         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 122         unsigned long    epp_st_outofmem;       /* # of out of mem requests */
 123         /*
 124          * pointers to pools, may be vmalloc'd
 125          */
 126         struct page    ***epp_pools;
 127 } page_pools;
 128
 129 /*
 130  * memory shrinker
 131  */
 132 static const int pools_shrinker_seeks = DEFAULT_SEEKS;
 133 static struct shrinker *pools_shrinker;
 134
 135
 136 /*
 137  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 138  */
 139 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 140 {
 141         spin_lock(&page_pools.epp_lock);
 142
 143         seq_printf(m, "physical pages:          %lu\n"
 144                    "pages per pool:          %lu\n"
 145                    "max pages:               %lu\n"
 146                    "max pools:               %u\n"
 147                    "total pages:             %lu\n"
 148                    "total free:              %lu\n"
 149                    "idle index:              %lu/100\n"
 150                    "last shrink:             %lds\n"
 151                    "last access:             %lds\n"
 152                    "max pages reached:       %lu\n"
 153                    "grows:                   %u\n"
 154                    "grows failure:           %u\n"
 155                    "shrinks:                 %u\n"
 156                    "cache access:            %lu\n"
 157                    "cache missing:           %lu\n"
 158                    "low free mark:           %lu\n"
 159                    "max waitqueue depth:     %u\n"
 160                    "max wait time:           "CFS_TIME_T"/%lu\n"
 161                    "out of mem:             %lu\n",
 162                    totalram_pages, PAGES_PER_POOL,
 163                    page_pools.epp_max_pages,
 164                    page_pools.epp_max_pools,
 165                    page_pools.epp_total_pages,
 166                    page_pools.epp_free_pages,
 167                    page_pools.epp_idle_idx,
 168                    cfs_time_current_sec() - page_pools.epp_last_shrink,
 169                    cfs_time_current_sec() - page_pools.epp_last_access,
 170                    page_pools.epp_st_max_pages,
 171                    page_pools.epp_st_grows,
 172                    page_pools.epp_st_grow_fails,
 173                    page_pools.epp_st_shrinks,
 174                    page_pools.epp_st_access,
 175                    page_pools.epp_st_missings,
 176                    page_pools.epp_st_lowfree,
 177                    page_pools.epp_st_max_wqlen,
 178                    page_pools.epp_st_max_wait,
 179                    msecs_to_jiffies(MSEC_PER_SEC),
 180                    page_pools.epp_st_outofmem);
 181
 182         spin_unlock(&page_pools.epp_lock);
 183         return 0;
 184 }
 185
 186 static void enc_pools_release_free_pages(long npages)
 187 {
 188         int     p_idx, g_idx;
 189         int     p_idx_max1, p_idx_max2;
 190
 191         LASSERT(npages > 0);
 192         LASSERT(npages <= page_pools.epp_free_pages);
 193         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 194
 195         /* max pool index before the release */
 196         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 197
 198         page_pools.epp_free_pages -= npages;
 199         page_pools.epp_total_pages -= npages;
 200
 201         /* max pool index after the release */
 202         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 203                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 204
 205         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 206         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 207         LASSERT(page_pools.epp_pools[p_idx]);
 208
 209         while (npages--) {
 210                 LASSERT(page_pools.epp_pools[p_idx]);
 211                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 212
 213                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 214                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 215
 216                 if (++g_idx == PAGES_PER_POOL) {
 217                         p_idx++;
 218                         g_idx = 0;
 219                 }
 220         }
 221
 222         /* free unused pools */
 223         while (p_idx_max1 < p_idx_max2) {
 224                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 225                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 226                 page_pools.epp_pools[p_idx_max2] = NULL;
 227                 p_idx_max2--;
 228         }
 229 }
 230
 231 /*
 232  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 233  */
 234 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 235                                             struct shrink_control *sc)
 236 {
 237         /*
 238          * if no pool access for a long time, we consider it's fully idle.
 239          * a little race here is fine.
 240          */
 241         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 242                      CACHE_QUIESCENT_PERIOD)) {
 243                 spin_lock(&page_pools.epp_lock);
 244                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 245                 spin_unlock(&page_pools.epp_lock);
 246         }
 247
 248         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 249         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 250                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 251 }
 252
 253 /*
 254  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 255  */
 256 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 257                                            struct shrink_control *sc)
 258 {
 259         spin_lock(&page_pools.epp_lock);
 260         sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 261                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 262         if (sc->nr_to_scan > 0) {
 263                 enc_pools_release_free_pages(sc->nr_to_scan);
 264                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 265                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 266
 267                 page_pools.epp_st_shrinks++;
 268                 page_pools.epp_last_shrink = cfs_time_current_sec();
 269         }
 270         spin_unlock(&page_pools.epp_lock);
 271
 272         /*
 273          * if no pool access for a long time, we consider it's fully idle.
 274          * a little race here is fine.
 275          */
 276         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 277                      CACHE_QUIESCENT_PERIOD)) {
 278                 spin_lock(&page_pools.epp_lock);
 279                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 280                 spin_unlock(&page_pools.epp_lock);
 281         }
 282
 283         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 284         return sc->nr_to_scan;
 285 }
 286
 287 #ifndef HAVE_SHRINKER_COUNT
 288 /*
 289  * could be called frequently for query (@nr_to_scan == 0).
 290  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 291  */
 292 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 293 {
 294         struct shrink_control scv = {
 295                 .nr_to_scan = shrink_param(sc, nr_to_scan),
 296                 .gfp_mask   = shrink_param(sc, gfp_mask)
 297         };
 298 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
 299         struct shrinker* shrinker = NULL;
 300 #endif
 301
 302         enc_pools_shrink_scan(shrinker, &scv);
 303
 304         return enc_pools_shrink_count(shrinker, &scv);
 305 }
 306
 307 #endif /* HAVE_SHRINKER_COUNT */
 308
 309 static inline
 310 int npages_to_npools(unsigned long npages)
 311 {
 312         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 313 }
 314
 315 /*
 316  * return how many pages cleaned up.
 317  */
 318 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 319 {
 320         unsigned long cleaned = 0;
 321         int           i, j;
 322
 323         for (i = 0; i < npools; i++) {
 324                 if (pools[i]) {
 325                         for (j = 0; j < PAGES_PER_POOL; j++) {
 326                                 if (pools[i][j]) {
 327                                         __free_page(pools[i][j]);
 328                                         cleaned++;
 329                                 }
 330                         }
 331                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 332                         pools[i] = NULL;
 333                 }
 334         }
 335
 336         return cleaned;
 337 }
 338
 339 /*
 340  * merge @npools pointed by @pools which contains @npages new pages
 341  * into current pools.
 342  *
 343  * we have options to avoid most memory copy with some tricks. but we choose
 344  * the simplest way to avoid complexity. It's not frequently called.
 345  */
 346 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 347 {
 348         int     freeslot;
 349         int     op_idx, np_idx, og_idx, ng_idx;
 350         int     cur_npools, end_npools;
 351
 352         LASSERT(npages > 0);
 353         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 354         LASSERT(npages_to_npools(npages) == npools);
 355         LASSERT(page_pools.epp_growing);
 356
 357         spin_lock(&page_pools.epp_lock);
 358
 359         /*
 360          * (1) fill all the free slots of current pools.
 361          */
 362         /* free slots are those left by rent pages, and the extra ones with
 363          * index >= total_pages, locate at the tail of last pool. */
 364         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 365         if (freeslot != 0)
 366                 freeslot = PAGES_PER_POOL - freeslot;
 367         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 368
 369         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 370         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 371         np_idx = npools - 1;
 372         ng_idx = (npages - 1) % PAGES_PER_POOL;
 373
 374         while (freeslot) {
 375                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 376                 LASSERT(pools[np_idx][ng_idx] != NULL);
 377
 378                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 379                 pools[np_idx][ng_idx] = NULL;
 380
 381                 freeslot--;
 382
 383                 if (++og_idx == PAGES_PER_POOL) {
 384                         op_idx++;
 385                         og_idx = 0;
 386                 }
 387                 if (--ng_idx < 0) {
 388                         if (np_idx == 0)
 389                                 break;
 390                         np_idx--;
 391                         ng_idx = PAGES_PER_POOL - 1;
 392                 }
 393         }
 394
 395         /*
 396          * (2) add pools if needed.
 397          */
 398         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 399                      PAGES_PER_POOL;
 400         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 401                      PAGES_PER_POOL;
 402         LASSERT(end_npools <= page_pools.epp_max_pools);
 403
 404         np_idx = 0;
 405         while (cur_npools < end_npools) {
 406                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 407                 LASSERT(np_idx < npools);
 408                 LASSERT(pools[np_idx] != NULL);
 409
 410                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 411                 pools[np_idx++] = NULL;
 412         }
 413
 414         page_pools.epp_total_pages += npages;
 415         page_pools.epp_free_pages += npages;
 416         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 417
 418         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 419                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 420
 421         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 422                page_pools.epp_total_pages);
 423
 424         spin_unlock(&page_pools.epp_lock);
 425 }
 426
 427 static int enc_pools_add_pages(int npages)
 428 {
 429         static DEFINE_MUTEX(add_pages_mutex);
 430         struct page   ***pools;
 431         int             npools, alloced = 0;
 432         int             i, j, rc = -ENOMEM;
 433
 434         if (npages < PTLRPC_MAX_BRW_PAGES)
 435                 npages = PTLRPC_MAX_BRW_PAGES;
 436
 437         mutex_lock(&add_pages_mutex);
 438
 439         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 440                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 441         LASSERT(npages > 0);
 442
 443         page_pools.epp_st_grows++;
 444
 445         npools = npages_to_npools(npages);
 446         OBD_ALLOC(pools, npools * sizeof(*pools));
 447         if (pools == NULL)
 448                 goto out;
 449
 450         for (i = 0; i < npools; i++) {
 451                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 452                 if (pools[i] == NULL)
 453                         goto out_pools;
 454
 455                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 456                         pools[i][j] = alloc_page(GFP_NOFS |
 457                                                  __GFP_HIGHMEM);
 458                         if (pools[i][j] == NULL)
 459                                 goto out_pools;
 460
 461                         alloced++;
 462                 }
 463         }
 464         LASSERT(alloced == npages);
 465
 466         enc_pools_insert(pools, npools, npages);
 467         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 468         rc = 0;
 469
 470 out_pools:
 471         enc_pools_cleanup(pools, npools);
 472         OBD_FREE(pools, npools * sizeof(*pools));
 473 out:
 474         if (rc) {
 475                 page_pools.epp_st_grow_fails++;
 476                 CERROR("Failed to allocate %d enc pages\n", npages);
 477         }
 478
 479         mutex_unlock(&add_pages_mutex);
 480         return rc;
 481 }
 482
 483 static inline void enc_pools_wakeup(void)
 484 {
 485         assert_spin_locked(&page_pools.epp_lock);
 486
 487         if (unlikely(page_pools.epp_waitqlen)) {
 488                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 489                 wake_up_all(&page_pools.epp_waitq);
 490         }
 491 }
 492
 493 static int enc_pools_should_grow(int page_needed, long now)
 494 {
 495         /* don't grow if someone else is growing the pools right now,
 496          * or the pools has reached its full capacity
 497          */
 498         if (page_pools.epp_growing ||
 499             page_pools.epp_total_pages == page_pools.epp_max_pages)
 500                 return 0;
 501
 502         /* if total pages is not enough, we need to grow */
 503         if (page_pools.epp_total_pages < page_needed)
 504                 return 1;
 505
 506         /*
 507          * we wanted to return 0 here if there was a shrink just happened
 508          * moment ago, but this may cause deadlock if both client and ost
 509          * live on single node.
 510          */
 511 #if 0
 512         if (now - page_pools.epp_last_shrink < 2)
 513                 return 0;
 514 #endif
 515
 516         /*
 517          * here we perhaps need consider other factors like wait queue
 518          * length, idle index, etc. ?
 519          */
 520
 521         /* grow the pools in any other cases */
 522         return 1;
 523 }
 524
 525 /*
 526  * Export the number of free pages in the pool
 527  */
 528 int get_free_pages_in_pool(void)
 529 {
 530         return page_pools.epp_free_pages;
 531 }
 532 EXPORT_SYMBOL(get_free_pages_in_pool);
 533
 534 /*
 535  * Let outside world know if enc_pool full capacity is reached
 536  */
 537 int pool_is_at_full_capacity(void)
 538 {
 539         return (page_pools.epp_total_pages == page_pools.epp_max_pages);
 540 }
 541 EXPORT_SYMBOL(pool_is_at_full_capacity);
 542
 543 /*
 544  * we allocate the requested pages atomically.
 545  */
 546 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 547 {
 548         wait_queue_t  waitlink;
 549         unsigned long   this_idle = -1;
 550         cfs_time_t      tick = 0;
 551         long            now;
 552         int             p_idx, g_idx;
 553         int             i;
 554
 555         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 556         LASSERT(desc->bd_iov_count > 0);
 557         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 558
 559         /* resent bulk, enc iov might have been allocated previously */
 560         if (GET_ENC_KIOV(desc) != NULL)
 561                 return 0;
 562
 563         OBD_ALLOC_LARGE(GET_ENC_KIOV(desc),
 564                   desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 565         if (GET_ENC_KIOV(desc) == NULL)
 566                 return -ENOMEM;
 567
 568         spin_lock(&page_pools.epp_lock);
 569
 570         page_pools.epp_st_access++;
 571 again:
 572         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 573                 if (tick == 0)
 574                         tick = cfs_time_current();
 575
 576                 now = cfs_time_current_sec();
 577
 578                 page_pools.epp_st_missings++;
 579                 page_pools.epp_pages_short += desc->bd_iov_count;
 580
 581                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 582                         page_pools.epp_growing = 1;
 583
 584                         spin_unlock(&page_pools.epp_lock);
 585                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 586                         spin_lock(&page_pools.epp_lock);
 587
 588                         page_pools.epp_growing = 0;
 589
 590                         enc_pools_wakeup();
 591                 } else {
 592                         if (page_pools.epp_growing) {
 593                                 if (++page_pools.epp_waitqlen >
 594                                     page_pools.epp_st_max_wqlen)
 595                                         page_pools.epp_st_max_wqlen =
 596                                                         page_pools.epp_waitqlen;
 597
 598                                 set_current_state(TASK_UNINTERRUPTIBLE);
 599                                 init_waitqueue_entry(&waitlink, current);
 600                                 add_wait_queue(&page_pools.epp_waitq,
 601                                                &waitlink);
 602
 603                                 spin_unlock(&page_pools.epp_lock);
 604                                 schedule();
 605                                 remove_wait_queue(&page_pools.epp_waitq,
 606                                                   &waitlink);
 607                                 LASSERT(page_pools.epp_waitqlen > 0);
 608                                 spin_lock(&page_pools.epp_lock);
 609                                 page_pools.epp_waitqlen--;
 610                         } else {
 611                                 /* ptlrpcd thread should not sleep in that case,
 612                                  * or deadlock may occur!
 613                                  * Instead, return -ENOMEM so that upper layers
 614                                  * will put request back in queue. */
 615                                 page_pools.epp_st_outofmem++;
 616                                 spin_unlock(&page_pools.epp_lock);
 617                                 OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 618                                                desc->bd_iov_count *
 619                                                 sizeof(*GET_ENC_KIOV(desc)));
 620                                 GET_ENC_KIOV(desc) = NULL;
 621                                 return -ENOMEM;
 622                         }
 623                 }
 624
 625                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 626                 page_pools.epp_pages_short -= desc->bd_iov_count;
 627
 628                 this_idle = 0;
 629                 goto again;
 630         }
 631
 632         /* record max wait time */
 633         if (unlikely(tick != 0)) {
 634                 tick = cfs_time_current() - tick;
 635                 if (tick > page_pools.epp_st_max_wait)
 636                         page_pools.epp_st_max_wait = tick;
 637         }
 638
 639         /* proceed with rest of allocation */
 640         page_pools.epp_free_pages -= desc->bd_iov_count;
 641
 642         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 643         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 644
 645         for (i = 0; i < desc->bd_iov_count; i++) {
 646                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 647                 BD_GET_ENC_KIOV(desc, i).kiov_page =
 648                        page_pools.epp_pools[p_idx][g_idx];
 649                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 650
 651                 if (++g_idx == PAGES_PER_POOL) {
 652                         p_idx++;
 653                         g_idx = 0;
 654                 }
 655         }
 656
 657         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 658                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 659
 660         /*
 661          * new idle index = (old * weight + new) / (weight + 1)
 662          */
 663         if (this_idle == -1) {
 664                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 665                             page_pools.epp_total_pages;
 666         }
 667         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 668                                    this_idle) /
 669                                   (IDLE_IDX_WEIGHT + 1);
 670
 671         page_pools.epp_last_access = cfs_time_current_sec();
 672
 673         spin_unlock(&page_pools.epp_lock);
 674         return 0;
 675 }
 676 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 677
 678 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 679 {
 680         int     p_idx, g_idx;
 681         int     i;
 682
 683         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 684
 685         if (GET_ENC_KIOV(desc) == NULL)
 686                 return;
 687
 688         LASSERT(desc->bd_iov_count > 0);
 689
 690         spin_lock(&page_pools.epp_lock);
 691
 692         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 693         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 694
 695         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 696                 page_pools.epp_total_pages);
 697         LASSERT(page_pools.epp_pools[p_idx]);
 698
 699         for (i = 0; i < desc->bd_iov_count; i++) {
 700                 LASSERT(BD_GET_ENC_KIOV(desc, i).kiov_page != NULL);
 701                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 702                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 703
 704                 page_pools.epp_pools[p_idx][g_idx] =
 705                         BD_GET_ENC_KIOV(desc, i).kiov_page;
 706
 707                 if (++g_idx == PAGES_PER_POOL) {
 708                         p_idx++;
 709                         g_idx = 0;
 710                 }
 711         }
 712
 713         page_pools.epp_free_pages += desc->bd_iov_count;
 714
 715         enc_pools_wakeup();
 716
 717         spin_unlock(&page_pools.epp_lock);
 718
 719         OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 720                  desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 721         GET_ENC_KIOV(desc) = NULL;
 722 }
 723
 724 /*
 725  * we don't do much stuff for add_user/del_user anymore, except adding some
 726  * initial pages in add_user() if current pools are empty, rest would be
 727  * handled by the pools's self-adaption.
 728  */
 729 int sptlrpc_enc_pool_add_user(void)
 730 {
 731         int     need_grow = 0;
 732
 733         spin_lock(&page_pools.epp_lock);
 734         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 735                 page_pools.epp_growing = 1;
 736                 need_grow = 1;
 737         }
 738         spin_unlock(&page_pools.epp_lock);
 739
 740         if (need_grow) {
 741                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 742                                     PTLRPC_MAX_BRW_PAGES);
 743
 744                 spin_lock(&page_pools.epp_lock);
 745                 page_pools.epp_growing = 0;
 746                 enc_pools_wakeup();
 747                 spin_unlock(&page_pools.epp_lock);
 748         }
 749         return 0;
 750 }
 751 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 752
 753 int sptlrpc_enc_pool_del_user(void)
 754 {
 755         return 0;
 756 }
 757 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 758
 759 static inline void enc_pools_alloc(void)
 760 {
 761         LASSERT(page_pools.epp_max_pools);
 762         OBD_ALLOC_LARGE(page_pools.epp_pools,
 763                         page_pools.epp_max_pools *
 764                         sizeof(*page_pools.epp_pools));
 765 }
 766
 767 static inline void enc_pools_free(void)
 768 {
 769         LASSERT(page_pools.epp_max_pools);
 770         LASSERT(page_pools.epp_pools);
 771
 772         OBD_FREE_LARGE(page_pools.epp_pools,
 773                        page_pools.epp_max_pools *
 774                        sizeof(*page_pools.epp_pools));
 775 }
 776
 777 int sptlrpc_enc_pool_init(void)
 778 {
 779         DEF_SHRINKER_VAR(shvar, enc_pools_shrink,
 780                          enc_pools_shrink_count, enc_pools_shrink_scan);
 781
 782         page_pools.epp_max_pages = totalram_pages / 8;
 783         if (enc_pool_max_memory_mb > 0 &&
 784             enc_pool_max_memory_mb <= (totalram_pages >> mult))
 785                 page_pools.epp_max_pages = enc_pool_max_memory_mb << mult;
 786
 787         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 788
 789         init_waitqueue_head(&page_pools.epp_waitq);
 790         page_pools.epp_waitqlen = 0;
 791         page_pools.epp_pages_short = 0;
 792
 793         page_pools.epp_growing = 0;
 794
 795         page_pools.epp_idle_idx = 0;
 796         page_pools.epp_last_shrink = cfs_time_current_sec();
 797         page_pools.epp_last_access = cfs_time_current_sec();
 798
 799         spin_lock_init(&page_pools.epp_lock);
 800         page_pools.epp_total_pages = 0;
 801         page_pools.epp_free_pages = 0;
 802
 803         page_pools.epp_st_max_pages = 0;
 804         page_pools.epp_st_grows = 0;
 805         page_pools.epp_st_grow_fails = 0;
 806         page_pools.epp_st_shrinks = 0;
 807         page_pools.epp_st_access = 0;
 808         page_pools.epp_st_missings = 0;
 809         page_pools.epp_st_lowfree = 0;
 810         page_pools.epp_st_max_wqlen = 0;
 811         page_pools.epp_st_max_wait = 0;
 812         page_pools.epp_st_outofmem = 0;
 813
 814         enc_pools_alloc();
 815         if (page_pools.epp_pools == NULL)
 816                 return -ENOMEM;
 817
 818         pools_shrinker = set_shrinker(pools_shrinker_seeks, &shvar);
 819         if (pools_shrinker == NULL) {
 820                 enc_pools_free();
 821                 return -ENOMEM;
 822         }
 823
 824         return 0;
 825 }
 826
 827 void sptlrpc_enc_pool_fini(void)
 828 {
 829         unsigned long cleaned, npools;
 830
 831         LASSERT(pools_shrinker);
 832         LASSERT(page_pools.epp_pools);
 833         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 834
 835         remove_shrinker(pools_shrinker);
 836
 837         npools = npages_to_npools(page_pools.epp_total_pages);
 838         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 839         LASSERT(cleaned == page_pools.epp_total_pages);
 840
 841         enc_pools_free();
 842
 843         if (page_pools.epp_st_access > 0) {
 844                 CDEBUG(D_SEC,
 845                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 846                        "access %lu, missing %lu, max qlen %u, max wait "
 847                        CFS_TIME_T"/%lu, out of mem %lu\n",
 848                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 849                        page_pools.epp_st_grow_fails,
 850                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 851                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 852                        page_pools.epp_st_max_wait,
 853                        msecs_to_jiffies(MSEC_PER_SEC),
 854                        page_pools.epp_st_outofmem);
 855         }
 856 }
 857
 858
 859 static int cfs_hash_alg_id[] = {
 860         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 861         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 862         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 863         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 864         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 865         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 866         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 867         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 868 };
 869 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 870 {
 871         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 872 }
 873
 874 __u8 sptlrpc_get_hash_alg(const char *algname)
 875 {
 876         return cfs_crypto_hash_alg(algname);
 877 }
 878
 879 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 880 {
 881         struct ptlrpc_bulk_sec_desc *bsd;
 882         int                          size = msg->lm_buflens[offset];
 883
 884         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 885         if (bsd == NULL) {
 886                 CERROR("Invalid bulk sec desc: size %d\n", size);
 887                 return -EINVAL;
 888         }
 889
 890         if (swabbed) {
 891                 __swab32s(&bsd->bsd_nob);
 892         }
 893
 894         if (unlikely(bsd->bsd_version != 0)) {
 895                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 896                 return -EPROTO;
 897         }
 898
 899         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 900                 CERROR("Invalid type %u\n", bsd->bsd_type);
 901                 return -EPROTO;
 902         }
 903
 904         /* FIXME more sanity check here */
 905
 906         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 907                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 908                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 909                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 910                 return -EPROTO;
 911         }
 912
 913         return 0;
 914 }
 915 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 916
 917 /*
 918  * Compute the checksum of an RPC buffer payload.  If the return \a buflen
 919  * is not large enough, truncate the result to fit so that it is possible
 920  * to use a hash function with a large hash space, but only use a part of
 921  * the resulting hash.
 922  */
 923 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 924                               void *buf, int buflen)
 925 {
 926         struct cfs_crypto_hash_desc     *hdesc;
 927         int                             hashsize;
 928         unsigned int                    bufsize;
 929         int                             i, err;
 930
 931         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 932         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 933         LASSERT(buflen >= 4);
 934
 935         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 936         if (IS_ERR(hdesc)) {
 937                 CERROR("Unable to initialize checksum hash %s\n",
 938                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 939                 return PTR_ERR(hdesc);
 940         }
 941
 942         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 943
 944         for (i = 0; i < desc->bd_iov_count; i++) {
 945                 cfs_crypto_hash_update_page(hdesc,
 946                                   BD_GET_KIOV(desc, i).kiov_page,
 947                                   BD_GET_KIOV(desc, i).kiov_offset &
 948                                               ~PAGE_MASK,
 949                                   BD_GET_KIOV(desc, i).kiov_len);
 950         }
 951
 952         if (hashsize > buflen) {
 953                 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
 954
 955                 bufsize = sizeof(hashbuf);
 956                 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
 957                          bufsize, hashsize);
 958                 err = cfs_crypto_hash_final(hdesc, hashbuf, &bufsize);
 959                 memcpy(buf, hashbuf, buflen);
 960         } else {
 961                 bufsize = buflen;
 962                 err = cfs_crypto_hash_final(hdesc, buf, &bufsize);
 963         }
 964
 965         return err;
 966 }